AI Power Demand Breaks Global Nuclear Supply Chains

The math is getting uncomfortable. AI data centers now consume more power per rack than small industrial facilities. A single NVIDIA GB200 NVL72 server rack draws up to 140 kilowatts, roughly the annual electricity consumption of 12 American households compressed into a cabinet the size of a refrigerator. Hyperscalers are buying these racks by the thousands, and they need that power continuously, around the clock, every day of the year. The result is a global energy transformation that is reshaping nuclear power economics, rewriting the rules for where AI infrastructure can be built, and exposing a geopolitical fault line that runs through Taiwan's aging electrical grid.

What Actually Happened

A Digitimes analysis published September 14, 2026 documented the accelerating collision between AI infrastructure demand and global electricity supply, focusing on how Taiwan's power grid is showing localized strain that reflects a worldwide structural pattern. Since 2024, hyperscalers have signed approximately 30 gigawatts of nuclear power purchase agreements with reactor operators across the United States, Europe, and Asia. Microsoft is refurbishing the former Three Mile Island Unit 1, renamed the Crane Clean Energy Center, under a 20-year PPA with Constellation Energy, targeting 835 megawatts of output by 2027. Amazon signed a deal with Talen Energy for 1,920 megawatts of nuclear capacity at the Susquehanna facility in Pennsylvania, running through 2042. Both deals represent a deliberate, multi-decade commitment to nuclear as the foundation of AI compute power.

The scale of this shift is historic. Data centers consumed approximately 23 gigawatts of electricity globally in 2023. By 2026, that figure has reached 42 gigawatts and is projected to double again by 2027 as AI workload intensity rises. Unlike conventional computing loads that peak during business hours, AI inference and training workloads run continuously. A frontier model training cluster may operate at 95% utilization for months without interruption. This 24/7 demand profile makes renewable energy with intermittent generation insufficient without massive storage infrastructure, and it is driving hyperscalers toward the only firm, carbon-free power source that operates on the same continuous schedule as their GPU clusters: nuclear. Informed Clearly reports that nearly half of U.S. AI data centers planned for 2026 are already delayed, creating a 7 gigawatt gap that is now bottlenecking $650 billion in hyperscaler capital expenditure.

Taiwan's situation illustrates the localized version of a global infrastructure problem. Taipower, the island's state-owned utility, received 79 AI data center power applications totaling approximately 4,758 megawatts as of late 2025. It approved only 40 of them, covering roughly 3,033 megawatts. The 39 rejected applications, covering 1,725 megawatts of demand, were turned away not because Taiwan lacks total electricity capacity but because specific regional substations and transmission corridors cannot handle the concentrated load that large AI campuses generate. Taiwan, which hosts TSMC's most advanced semiconductor fabs and is central to the global AI chip supply chain, is now actively considering restarting idle nuclear reactors to address this localized grid strain, representing a potential reversal of its previous phase-out policy driven by post-Fukushima public sentiment. According to the International Atomic Energy Agency, this pattern of AI-driven nuclear reconsideration is appearing across multiple countries simultaneously.

Why This Matters More Than People Think

The hyperscalers are not accidentally creating a nuclear renaissance. They are deliberately engineering one because no alternative works at the scale and reliability profile their AI operations require. Solar and wind generate power intermittently, and the cost of battery storage sufficient to guarantee 24/7 power for a 500-megawatt AI campus over a winter week of low solar irradiance and wind lulls is prohibitive. Natural gas provides firm power but carbon net-zero commitments make it politically and legally difficult for companies with published climate pledges to rely on it as a primary source for new facilities. Nuclear is the only technology that produces dense, reliable, carbon-free electricity on the continuous schedule that AI infrastructure demands. When Microsoft, Amazon, and Google collectively sign 30 gigawatts of nuclear PPAs, they are not hedging a climate bet. They are solving an engineering constraint that has no other viable solution at this scale.

The consequences extend far beyond energy markets. The 7-gigawatt gap in delayed U.S. data center construction is not a minor inconvenience. It represents a binding constraint on how fast AI compute capacity can grow, regardless of how many NVIDIA Blackwell or GB200 chips are manufactured and shipped. Data centers that cannot get power interconnect approvals cannot break ground. Those that cannot break ground cannot host the AI training and inference capacity that determines competitive position in the model race. That means compute supply growth is now constrained not primarily by chip availability, which has improved substantially with TSMC's Arizona expansion ramping, but by the slowest variable in the entire system: electrical infrastructure that takes 5 to 10 years to plan, permit, and deploy from initial application to energization. The AI race has become, in a structurally important sense, an electricity procurement race.

The geopolitical dimension is the piece most analysts are underweighting when they assess the long-term trajectory of AI infrastructure. Taiwan's semiconductor industry is already the focal point of U.S.-China technology competition, and its AI data center energy challenges are compounding the strategic complexity of the island's position. Taiwan has limited land area, a constrained transmission grid with regional substation bottlenecks in the industrial zones where fabs are concentrated, affecting over 1,700 MW of rejected power applications, and a population that has historically been ambivalent about nuclear power after Fukushima. If Taiwan restarts idle reactors to accommodate AI infrastructure demand from companies like TSMC and its hyperscaler customers, it will be making a domestic political commitment that is simultaneously a bet on sustained stability in the Taiwan Strait. An island that becomes more energy-dependent on nuclear facilities to sustain AI-critical semiconductor manufacturing becomes harder, not easier, to transition or protect in a geopolitical disruption scenario.

The Competitive Landscape

The competition in nuclear power access is no longer between utilities or governments. It is between hyperscalers racing to lock up the finite supply of existing and planned reactor capacity before competitors do. Microsoft's Three Mile Island PPA effectively took that plant's entire output off the market for every other buyer for 20 years. Amazon's Susquehanna deal similarly monopolizes the most reliable nuclear supply in the U.S. mid-Atlantic region. Google has secured undisclosed nuclear capacity in the Midwest through deals that have not been fully disclosed. This race to sign PPAs is functioning like a land grab, where first movers secure the best sites and the longest contract durations, and latecomers face longer lead times, higher premia, and less reliable interconnect positions.

The competition extends into small modular reactors, where Kairos Power, NuScale, X-energy, and Oklo are all receiving early commitments and conditional agreements from hyperscalers who cannot wait the decade it typically takes to refurbish and restart existing light water reactors. Microsoft invested in Helion Energy's fusion program as a longer-term hedge beyond fission. The challenge is that no SMR design has achieved commercial electricity generation at scale in the United States, and the regulatory, construction, and commissioning timelines for new nuclear plants in Western regulatory environments remain measured in years from initial licensing to first electricity. The market is pricing nuclear capacity as if SMRs will deliver on the schedule their developers have published, which is optimistic given the construction and budget history of nuclear projects in the U.S. and European regulatory contexts of the last two decades.

The historical parallel that best captures this dynamic is the late 19th-century railroad expansion, when American industrial capacity was constrained not by raw material availability but by the speed at which rail infrastructure could connect mines, mills, and ports. Companies and commodities that controlled key rail routes controlled industrial throughput for entire regions. The railroads that built first set the gauge standards, the right-of-way precedents, and the rate structures that shaped American industry for a generation. Today, companies that secure nuclear power purchase agreements and data center campuses near reliable grid interconnect points are not just managing electricity costs. They are controlling the physical infrastructure that determines AI compute throughput for the next decade, and the first movers are establishing the advantages that compound most durably in capital-intensive infrastructure industries.

Hidden Insight: The Supply Chain Bottleneck No One Is Modeling

The Taiwan grid story reveals something structurally uncomfortable about the way AI infrastructure investment is being planned globally. Taiwan produces the world's most advanced semiconductors, yet it cannot reliably power the data centers required to deploy those semiconductors at the scale the AI industry is demanding. This is not a planning failure or a governance failure in the usual sense. It is a geographic and temporal mismatch: chip design timelines run in product cycles of 18 to 24 months, while electrical infrastructure investment cycles run in decades. TSMC's N2 process node fabs are coming online now, but the power grid supplying the industrial zones where those fabs operate has not materially expanded since investment decisions made in the early 2010s, when AI infrastructure demand was not part of any power planner's 20-year load forecast.

The supply chain tightening that the Digitimes analysis documented extends far beyond Taiwan. Global nuclear power construction requires specialized steel forgings for pressure vessels, precision pumps for primary coolant loops, reactor vessel components that can only be manufactured at a handful of facilities worldwide, and trained nuclear engineers whose pipeline was severely contracted during the post-Chernobyl and post-Fukushima decades when nuclear construction nearly stopped in Western markets. A worldwide rush to build new reactor capacity is already competing for the same forgings, the same concrete formwork, the same regulatory engineering expertise, and the same specialized construction labor. The IAEA has flagged labor and manufacturing supply constraints as primary bottlenecks for new nuclear construction globally. When hyperscalers simultaneously commit to 30 gigawatts of nuclear capacity, they collectively create a procurement problem that no individual company's engineering team can solve through faster execution.

There is a second-order effect on AI model economics that the market is not pricing correctly. If electricity constraints slow compute capacity growth from its historical trajectory, AI inference costs will not fall as quickly as the GPU performance curve alone would project. The standard assumption across AI business model planning is that compute gets cheaper by approximately 30-40% per year, following a pattern similar to semiconductor learning curves. That projection assumes unconstrained power capacity at each generation of hardware. If 7 gigawatts of planned data center capacity is delayed by 18 to 24 months, the effective compute supply shortfall compounds during the delay period. The AI-native applications and business models being designed today for a world of cheap, abundant inference may find themselves operating in a world where electricity scarcity creates an unexpected and durable price floor that no amount of chip efficiency improvement fully overcomes.

The bear case for the nuclear renaissance is real and deserves explicit acknowledgment. Critics argue that hyperscalers are overpaying for nuclear capacity based on projections of AI demand growth that may materialize on a different timeline or at a different geographic distribution than the current wave of PPA signings assumes. The history of infrastructure investment cycles includes prominent examples where capacity was built for a demand wave that arrived later than expected, was more concentrated than forecast, or was fundamentally different in character from what drove the original investment case. If energy-efficient AI inference techniques, including model distillation, speculative decoding, and hardware-software co-optimization, reduce power consumption per AI query by 60-70% over the next three years, the 30 gigawatts of nuclear PPAs being signed in 2025 and 2026 could prove to be expensive overcorrection. Signing 20-year PPAs for power that proves excessive is not a trivial commitment for any organization's balance sheet.

What to Watch Next

Watch Constellation Energy's Three Mile Island restart timeline over the next 30 to 90 days. The plant is targeting its return to service in 2027, but refurbishment projects of this regulatory and engineering complexity have historically run over schedule and over budget in the U.S. nuclear industry. If the restart faces Nuclear Regulatory Commission review delays, unexpected infrastructure findings during the refurbishment, or labor disputes at the construction site, it will signal that the first wave of nuclear power commitments from hyperscalers carries more execution risk than the power purchase agreement terms suggest. Constellation's quarterly earnings calls through 2026 and into early 2027 will be the primary public data source for tracking restart progress against schedule.

Taiwan's nuclear policy response is the 180-day indicator worth tracking most carefully. If the Taiwanese government formally approves the restart of any of its mothballed nuclear units before the end of 2026, it will mark the first explicit documented case of a government reversing a nuclear phase-out policy specifically in direct response to AI infrastructure demand. That decision would establish a precedent that other governments with idle nuclear capacity, including Germany, Belgium, and Japan, are watching closely. A domino effect of governments re-embracing nuclear to power AI infrastructure and semiconductor manufacturing could reshape the global energy market for the next three decades, affecting climate policy, energy security strategy, and electricity pricing across entire continents.

Within 30 days, watch for whether the IAEA or the International Energy Agency publishes updated guidance specifically addressing nuclear supply chain constraints relative to the pace of hyperscaler PPA commitments. Both agencies have been tracking the acceleration in nuclear power purchase agreements and have flagged concern about the possibility of a structural mismatch between global commitments and available manufacturing capacity for reactor components. If either agency issues a formal assessment warning that hyperscalers have collectively signed more nuclear capacity agreements than the global nuclear manufacturing and construction industry can fulfill within the announced timelines, it will be the clearest independent signal that the current pace of AI-driven nuclear investment is outrunning physical supply chain realities.

When every hyperscaler signs a 20-year nuclear deal simultaneously, they have not solved the AI power problem. They have just made it more expensive for everyone else to solve theirs.

---